Agent Evaluation

Evaluate an agent at three levels — build a golden trajectory dataset, match trajectories with agentevals and LLM-as-judge, measure tool-calling accuracy, and gate it in CI with LangSmith and Phoenix.